[MooreToCore] More extract op lowerings

[maerhart]

We need to inset runtime checks for out-of-bound accesses because the semantics of the core dialects don't match the ones of SystemVerilog. I left this for a future PR since there are a few interesting cases to consider and it would increase the size of this PR considerably.

[hailongSun2000]

It makes sense! We ever discussed bits slice. @fabianschuiki thinks we can refactor moore.extract to extract, array_get, and array_slice. But I think hw.array_get cannot handle the unpacked array type, so maybe we can keep moore.extract to extract some bits from the vector, N-dimension packed/unpacked array at the moore level. It's also a great option that distinguishes them at the core level. LGTM 🥳 !

[hailongSun2000]

I viewed the IEEE 1800-2017 § 7.4.6 "Indexing and slicing of arrays", if we access the index(out of bound), we can return a value based on Table 7-1.

[maerhart]

Yes, it might be worth refactoring the extract operations a bit. I'm also wondering if it's worth having a static extract operation for arrays if we already have the dynamic one and could just pass a constant value. I also agree that using the same operation for array_get and array_slice is strange.

Thanks for looking it up in the standard! The problem we have here is that SV defines exactly that we need to return 0 or X, but CIRCT Core allows any value within the type constraint. So we need to insert a comb.muxthat returns 0/X when out of bounds.

[hailongSun2000]

I think that's worth it. It's a great idea to insert a comb.mux.

[fabianschuiki]

Very nice 🥳!